Two AI agents inventing their own language sounds like science fiction — it is increasingly a research staple.
Researchers ran LLM agents through a classic Lewis signaling game, where a sender and receiver must coordinate on a shared code using only their interaction history. They tested five memory architectures across different channel capacities — essentially, how many distinct signals agents can use. The headline finding: memory architecture outweighs channel capacity. Agents given a persistent private notebook reached a coordination score of 0.867 at a channel capacity of 25, the most reliable result in the study. Stateless agents, by contrast, peaked at moderate capacity and fell apart as the vocabulary grew beyond what a rolling context window could track.
The practical implication is pointed. Current agent design debates tend to fixate on context window size and token limits — proxies for capacity. This work suggests that externalizing learned conventions, giving agents something like a scratchpad that persists across rounds, does more to stabilize coordination than simply widening the channel. The notebook lets agents stop re-deriving the same codes each round and start building on them.
The study also punctures a tidier theory: an information bottleneck argument predicted the optimal channel capacity would equal the number of objects in the task, but that threshold turned out to be a fragility point, not a sweet spot. More headroom, not less, kept coordination stable — which is a useful corrective to anyone designing agent communication protocols on the assumption that constraint breeds clarity.